OpenAI Highlights New GPU Networking Protocol and Partnerships for Frontier Model Training

According to a recent LinkedIn post from OpenAI, the company is emphasizing the importance of high-performance networking for training frontier AI models at scale. The post describes how GPU clusters can be bottlenecked by slow data movement or failures in links and switches, which can idle expensive hardware and disrupt long training runs.

The company’s LinkedIn post highlights a partnership with AMD, Broadcom, Intel, Microsoft, and NVIDIA to develop Multipath Reliable Connection (MRC), an open protocol designed for large GPU training clusters. MRC is described as distributing traffic across hundreds of network paths and routing around failures quickly to improve utilization of compute resources.

According to the post, MRC is already deployed on OpenAI’s largest supercomputers, including its Oracle Cloud Infrastructure site in Abilene, Texas, and Microsoft’s Fairwater systems. The specification has reportedly been released through the Open Compute Project Foundation, suggesting an intention to encourage broader industry adoption and ecosystem development.

For investors, the post suggests OpenAI is investing not only in models but also in underlying infrastructure, which could enhance training efficiency and reduce per-model compute costs over time. The collaboration with major chipmakers and cloud providers may reinforce OpenAI’s position within a strategic hardware and cloud ecosystem, potentially improving its access to cutting-edge compute at scale.

If widely adopted, MRC could influence standards for GPU networking in large AI clusters, with implications for capital spending and technical roadmaps across hyperscalers and semiconductor vendors. The focus on open protocols and partner alignment may also mitigate vendor lock-in risks for OpenAI while deepening its integration with key infrastructure partners.